Medium scale approach (MSA) for improved assessment of ...adn.biol.umontreal.ca/~numericalecology/Reprints/Clua_et_al_JEMBE.pdf · Medium scale approach (MSA) for improved assessment

y and Ecology 333 (2006) 219–230www.elsevier.com/locate/jembe

Journal of Experimental Marine Biolog

Medium scale approach (MSA) for improved assessmentof coral reef fish habitat

Eric Clua a,b,⁎, Pierre Legendre c, Laurent Vigliola b, Franck Magron b,Michel Kulbicki d, Sébastien Sarramegna e, Pierre Labrosse f, René Galzin a

a École Pratique des Hautes Études, UMR 8046 EPHE-CNRS, Université de Perpignan, 54bis Avenue Paul Alduy, 66860 Perpignan, Franceb Secretariat of the Pacific Community, BP D5, 98848 Nouméa, New Caledonia

c Département de Sciences Biologiques, Université de Montréal, C.P. 6128, Montréal, Québec H3C 3J7, Canadad Institut de la Recherche pour le Développement, BP A5, 78848 Nouméa, New Caledoniae Falconbridge, 9, rue d'Austerlitz, BP MGA08, 98802 Nouméa Cedex, New Caledonia

f BP 1386, Dakar, Sénégal

Received 29 August 2005; received in revised form 20 December 2005; accepted 20 December 2005

Abstract

Habitat characteristics play a critical role in structuring reef fish communities subjected to fishing pressure. The line intercepttransect (LIT) method provides an accurate quantitative description of the habitat, but in a very narrow corridor less than 1 m wide.Such a scale is poorly adapted to the wide-ranging species that account for a significant part of these assemblages. We developed aneasy-to-use medium scale approach (MSA), based on a semi-quantitative description of 20 quadrats of 25 m2 (500 m2 in total). Wethen simulated virtual reef landscapes of different complexities in a computer, on which we computed MSA using differentmethods of calculation. These simulations allowed us to select the best method of calculation, obtaining quantitative estimates withacceptable accuracy (comparison with the original simulated landscapes: R2 ranging from 0.986 to 0.997); they also showed thatMSA is a more efficient estimator than LIT, generating percentage coverage estimates that are less variable. A mensurativeexperiment based on thirty 50-m transects, conducted by three teams of two divers, was used to empirically compare the twoestimators and assess their ability to predict fish–habitat relationships. Three-factor multivariate ANOVAs (Teams, Reef, Methods)revealed again that LIT produced habitat composition estimates that were more variable than MSA. Canonical analyses conductedon fish biomass data successively aggregated by mobility patterns, trophic groups, and size classes, showed the higher predictivepower of MSA habitat data over LIT. The MSA enriches the toolbox of methods available for reef habitat description atintermediate scale (b1000 m2), between the scale where LIT is appropriate (b100 m2) and the landscape approach (N1000 m2).© 2005 Elsevier B.V. All rights reserved.

Keywords: Biomass; Fish–habitat relationships; Line intercept transect; Medium scale approach; Monte Carlo simulations; Reef habitat description;Reef surveys; Variance of estimator

⁎ Corresponding author. Secretariat of the Pacific community, BP D5, 98848 Nouméa, New Caledonia. Tel./fax: +687 265471.E-mail addresses: [email protected] (E. Clua), [email protected] (P. Legendre), [email protected] (L. Vigliola), [email protected]

(F. Magron), [email protected] (M. Kulbicki), [email protected] (S. Sarramegna), [email protected] (P. Labrosse),[email protected] (R. Galzin).

0022-0981/$ - see front matter © 2005 Elsevier B.V. All rights reserved.doi:10.1016/j.jembe.2005.12.010

220 E. Clua et al. / Journal of Experimental Marine Biology and Ecology 333 (2006) 219–230

1. Introduction

Many studies have been conducted to identify therelationships between reef fishes and their habitat.Most of them consider the whole set of species formingreef fish assemblages, including the small and gre-garious species, and rely on a fine-scale approach suchas the widely used line intercept transect method (LIT;English et al., 1997), to estimate habitat characteristics.This technique provides an accurate quantitativedescription of the habitat, but in a very narrow corridor(less than 1 m wide) and is time-consuming. Long et al.(2004) have shown that new techniques based on visualestimates of percentage cover of benthos and substra-tum (such as the Reef Resource Inventory, RRI) pro-vide comparable sampling accuracy with a relative costefficiency at least three times that of LIT. However,RRI is conducted along two 20-m plotless strip-transects, which represent a working scale comparableto that of LIT. Therefore, if RRI increases cost ef-ficiency during intensive field surveys, it does notaddress the problem of correspondence of scale andmay still be poorly adapted to the wide-ranging spe-cies, which account for a significant portion of reef fishassemblages.

Many authors have shown the importance of localstructuring factors of reef fish assemblages such ashabitat complexity (Grigg, 1994; Caley and St John,1996; Beukers and Jones, 1998), shelter availability(Connell and Kingsford, 1998; Friedlander and Parrish,1998a) or habitat rugosity (Luckhurst and Luckhurst,1978; McClanahan, 1994). These factors are often cor-related with one another, each one contributing to thegeneral and complex concept of “heterogeneity” asdescribed by Kolasa and Rollo (1991). These authorsinsist on the importance of estimating environmentalheterogeneity at the scale at which the organisms per-ceive it. “Functional heterogeneity” is the heterogeneitythat an organism perceives and responds to. It maydiffer from heterogeneity estimated using arbitrary eco-logical measures, and a discrepancy between the scalesof collection of the fish and habitat data may producebiased results (Jones and Syms, 1998; González-Gándara et al., 1999). Since many of the edible fishspecies have a much greater range of activity than thenarrow corridor assessed by LIT or RRI, this suggeststhat the scale of description of the habitat should beincreased to make it closer to that of the fish. Thisconcern is also the foundation of survey methods,such as distance sampling which has been adaptedto underwater visual censuses (UVC), in which thesurveyor counts fish over several metres (usually up

to 10 m) on either side of a transect (Labrosse et al.,2003).

We know that a statistical estimator A is moreefficient than an estimator B if, for equal sample sizes(n), the variance of A is smaller than that of B (Mikulski,1982). The size of a sampling unit has a critical effect onour perception of ecological phenomena; it influencesthe variance and correlation structure estimates. Usinggeostatistical theory, Bellehumeur et al. (1997) showedthat, as the size of the sampling units increases, thevariance and proportion of noise in the observed datadecreases. Based on these evidences and the fact that90% of the coral reef fish have life territories smallerthan 20 m2 (Galzin and Harmelin-Vivien, 2000), wedeveloped a medium scale approach (MSA) for habitatassessment (on 20 quadrats of 5×5 m) for the specificpurpose of better assessing habitat–fish assemblagerelationships when studying certain stocks of reef fishthat are of interest for coastal reef fisheries. We couldthus expect a MSA estimator, which is based on largersampling units, to display less variance and, thus, bemore efficient than a LIT estimator.

However, surveying at broader habitat scale maybe more time-consuming if a quantitative approachis maintained. This, in turn, could make the imple-mentation of a MSA approach difficult during intensivefield surveys. A semi-quantitative approach, alreadypromoted for habitat assessment of small quadrats of1×1 m (English et al., 1997), appeared more suitable,but it needed to be validated at broader scale. If it couldbe shown to be theoretically reliable, it would beinteresting to compare its efficiency with LIT forassessing fish–habitat relationships during a fieldsurvey.

In this paper, we first describe the method that wedeveloped to associate suitable scale, speed, and fieldeffectiveness for the description of the habitat of coralreef fish assemblages targeted by fishing. We thenanswer the following questions:

1) Based on numerical simulations, is the semi-quantitative approach acceptable for describing thehabitat at the proposed scale?

2) Does the MSA estimator generate less variance thanthe LIT estimator, making the former a more efficienttool for surveying complex coral habitats?

3) Using real data on reef habitat and fish stocks,collected during a mensurative experiment involvingboth estimators and taking into consideration ob-server bias during their implementation, does thenew MSA method lead to a better assessment of thefish–habitat relationships?

221E. Clua et al. / Journal of Experimental Marine Biology and Ecology 333 (2006) 219–230

2. Materials and methods

2.1. Habitat description

The new method of habitat description (MSA) wasdeveloped as a complement to the distance-samplingunderwater visual census (UVC) method for fish surveys,first developed by Kulbicki and Sarramegna (1999) andfully described by Labrosse et al. (2003). Ten 5×5 mquadrats are delimited on each side of a 50-m transectmaterialized on the seafloor by a measuring tape, for atotal of 20 quadrats per transect (Fig. 1). The 5-m scalewas imposed by the difficulty for an unmoving diver ofproperly describing the habitat over an area larger than25 m2. In each quadrat, depth was measured in the centreof the quadrat using a dive computer. Sixteen substratecomponents, totalling 100% covering, were recorded ifpresent. They were divided in two groups. The firstgroup contains 9 abiotic components: 1. mud (sedimentparticlesb0.1 mm), 2. sand and gravel (0.1 mmbhardparticlesb30mm), 3. small boulders (diameterb30 cm),4. big boulders (diameterb1 m), 5. rock (massive min-

Fig. 1. Graphical representation of a 50-m transect divided in 20 quadrats of 2made available on the data sheet in order to ease the evaluation of the main comtransect is mainly composed of sand (coefficient 4), then massive coral (coevalues of interval percentages for these three coefficients is z100%. Depth

erals) and eroded dead coral (carbonated edifices thathave lost their coral colony shape), 6. slab (flat rock withno relief), 7. dead coral debris (carbonated structures ofheterogeneous sizes, broken and removed from theiroriginal locations), 8. branching dead coral (deadcarbonated edifices that are still in place and retain ageneral branching coral shape), and 9. massive deadcoral (same, but massive shape). The second groupcontains 7 “live coral” shapes (English et al., 1997):1. encrusting, 2. massive, 3. digitate and submassive,4. foliose, 5. table, 6. small branches (segmentsb10 cm),and 7. large branches (segmentsN10 cm). Eachcomponent was quickly estimated using a semi-quantitative scale (SQS): 0 (0%), 1 (1–10%), 2 (11–30%), 3 (31–50%), 4 (51–75%) and 5 (76–100%)(English et al., 1997) (Fig. 1). The coverage coefficientswere allocated based on a “single layer approach” whereonly the visible surfaces were taken into consideration,all components being projected vertically. To make surethat the SQS was properly used, operators were asked toverify that after coding the habitat components, the sumof the highest values of interval percentages (described

5 m2 each for habitat description. The SQS (semi-quantitative scale) isponents. In the example, the last quadrat (No. 10) of the left side of thefficient 2) and a small boulder (coefficient 1). The sum of the highestis measured at the centre of a quadrat.


above) allowed to reach 100% and the sum of the lowestvalues of these intervals did not exceed 100%. Inaddition to the first layer of abiotic and live coralsubstrates, a second layer made of bleached coral, softcoral, anemones, sponges, macro-algae (Sargassum sp.,Lobophora sp., Turbinaria sp., Caulerpa sp., Halimedasp.), encrusting algae and seagrass (phanerogams) wasrecorded using the same semi-quantitative scale as forlive coral. Due to their high frequency, micro-algae(turf) were recorded using a semi-quantitative scalefrom 1 to 5 which took into consideration both theirsurface and volume.

2.2. Habitat simulations to validate SQS

We performed Monte Carlo simulations to comparevarious ways of computing overall estimates of habitatcomposition along transects, from habitat componentsestimated in quadrats using a semi-quantitative scale.These simulations involved transects, divided into 10quadrats, placed at random in landscapes with differenthabitat compositions. These simulations were notspatially explicit, meaning that the transects were notgeographically located in an area on which the habitatcomponents would have been previously mapped.

The method was the following. (1) The followingvalues of intra-transect coefficients of variation (ITCV)were chosen for the simulations: 0.1, 0.5, 1, 10, and 100.(2) For each of 500 simulations (5 different ITCV values,each with 100 random transects), we generated a newlandscape and a random transect in that landscape. First,we generated a vector of percentage coverage values,Cr(i), for the 10 habitat components i in the landscape;these values were drawn at random from a uniformdistribution and normalized to a sum of 100%. For eachhabitat component, we calculated the intra-transectstandard deviation of the landscape, ITSDi=ITCV×Cri.Thenwe generated the values present in the 10 quadrats (j)of a random transect as follows: for each habitat com-ponent, we drew 10 values at random from a normaldistribution with mean Cri and standard deviation ITSDi.For each quadrat, the values were normalized to a sum of100% over the 10 habitat components and then trans-formed to semi-quantitative notation from 0 to 5, fol-lowing the semi-quantitative scale (SQS) described in theprevious section. These values were assembled in a worktable with rows (i) corresponding to the habitat compo-nents and columns (j) corresponding to the quadrats of thetransect. (3) Calculation methods 1 to 4 (Table 1) wereapplied to that table to estimate the relative coveragevalues of the 10 habitat components. Finally, for eachcalculation method, a coefficient of determination (R2)

was computed, over 1000 pairs of values (10 habi-tats×100 transects), between the estimated relativecoverage values and the values in vector Cr for thesimulated landscape. The calculation was repeated using5000 pairs of values (5 ITCV values×10 habitats×100transects) to obtain a global R2 for each calculationmethod. The best calculation method to estimate habitatcomposition from semi-quantitative data has been used inthe next series of simulations and during the field survey.

2.3. Comparison of the MSA and LIT estimators bysimulation

Using the PERL programming language, a set of fivevirtual reef landscapes was set up, each comprising 10randomly generated components, with the constraintthat the total surface of each landscape was 100%. Thesesimulations were spatially explicit, meaning that thetransects were geographically located in an area onwhich the habitat components had been mapped.

In the present study, we consider that the complexity ofa habitat is linked to the number of components and theirrelative surfaces. As in the previous section, the morebalanced the components are on a given total surface, themore complex the habitat is. We estimated complexitythrough a coefficient of variation (ratio between standarddeviation and mean of the surface habitat components),which is negatively correlated with complexity (Table 2).

For each landscape, 2500 transects were randomlygenerated and, for each one, the habitat componentswere estimated by LIT and MSA. For each transect andestimation method, the total squared estimation errorwas computed as the sum, over the 10 components, ofthe squared differences between the estimated and realhabitat component percentages. The mean, over all2500 transects of a landscape, was used to represent theactual percentage of a component for the landscape inthe computation of the sum of squared differences. Themean of the squared errors, over the 2500 transects,estimates the variance due to the method. It was com-puted for each landscape and method (LIT and MSA).

2.4. Comparison of the MSA and LIT estimators duringa mensurative experiment

A total of 30 transects were surveyed on the south-west coast of New Caledonia (20°57′S to 21°14′S,164°32′E to 164°46′E); the survey was structured as amensurative experiment (sensu Hurlbert, 1984). Thetransects were located in two different biotopes, themiddle reef area and the inner barrier reef, in front of thetown of Koné. Depth was less than 5 m for all transects.

Table 1Four methods of calculation of habitat component coverage percentages, compared in the habitat simulations to validate the semi-quantitative scale(SQS)

Calculation methods Formulae

1. Sum the SQS scores in each row (habitat component i). Divide these sums by thegrand total to normalize the results column. Cei ¼

Pj SQS scoresði; jÞ

Pi;j SQS scoresði; jÞ

2. Normalize the SQS scores in each quadrat (column j) by dividing each SQS scoreby the column sum. Then, sum the normalized values per row (habitat componenti) and normalize these sums as above.

Cei ¼P

jSQS scoresði;jÞPijj SQS scoresði;jÞ

Pi;j

SQS scoresði;jÞPijj SQS scoresði;jÞ

3. Replace each SQS score by the median percentage coverage value (SQS of 0, 1, 2,3, 4 and 5 are replaced by 0%, 5.5%, 20.5%, 40.5%, 63%, and 88%, respectively).Apply the formula of method 1 to the transformed SQS scores.

Cei ¼P

j Transformed SQS scoresði; jÞP

i;j Transformed SQS scoresði; jÞ

4. Replace each SQS score by the median percentage coverage value, as in method 3.Normalize each column (quadrat j) as in method 2. Then, calculate the mean valueper row i (habitat component).

Cei ¼ 1No: of quadrats

X

j

Transformed SQS scoresði; jÞPijj Transformed SQS scoresði; jÞ

Hab

itat

i

...

1

1Score(1,1)

...Score(i,1)

Total ∑i Score(i,1)

...

...

...

...

jScore(1, j)

...

Score(i, j)

∑i Score(i, j)

Total

∑j Score(i, j)

...

∑j Score(i, j)

∑i,j Score(i, j)

Quadrat


The transects were surveyed by 3 teams of 2 divers each.Each team had 5 transects to survey in each reef area.Underwater visual censuses (UVC) along 50-m trans-ects were carried out to estimate the fish stocks. Thismethod, which is an adaptation to the underwater en-vironment (Kulbicki and Sarramegna, 1999; Kulbicki etal., 2000) of the transect method of Buckland et al.(2001) for assessment of animal densities, has been fullydescribed by Labrosse et al. (2003). It allows the in-clusion of the mobile and shy species in surveys; most ofthese species are of fishing interest. Two divers werepulling a tape while counting fish on either side of theline and recording the perpendicular distance betweenthe fish and the transect line. The total length of each

Table 2Distribution of ten components (columns 1–10) randomly generated in five

Components ⇒ 1 2 3 4 5

Landscape 1 18.5 8.3 9.7 7.1 8.2Landscape 2 21.0 7.0 7.6 7.5 7.4Landscape 3 25.0 7.7 8.9 7.8 8.9Landscape 4 59.8 4.8 4.9 4.2 4.4Landscape 5 75.1 3.3 2.6 2.3 3.0

The component surfaces are given as percentages; row sums are 100. The coedefinition of complexity, landscapes are listed in order of decreasing comple

fish was estimated using 1-cm classes from 4 to 10 cm,2-cm classes from 10 to 30 cm, 5-cm classes from 30 to60 cm, and 10-cm classes above 60 cm. Biomass es-timates were calculated based on the total length–massrelationships (Letourneur et al., 1998) and densitiesestimated from the fish counts, and mean perpendiculardistance from the transect (Labrosse et al., 2003). Whenboth divers had completed the fish census, habitatvariables were recorded on the way back. For habitatdescription, the divers used successively (1) the MSAmethod described in this paper and (2) the LIT methoddescribed by English et al. (1997). The divers had neverused either method in the field before this survey butthey were familiar with the reef environment and had

virtual reef landscapes

6 7 8 9 10 CV

12.2 9.7 9.9 8.4 7.9 0.39.3 12.5 9.8 6.1 11.8 0.47.0 7.5 8.7 10.7 7.7 0.53.6 5.3 4.7 4.5 3.8 1.72.4 2.7 2.6 3.3 2.6 2.2

fficient of variation (CV) was calculated for each landscape. Given ourxity.

Table 3Coefficients of determination (R2) of the linear regressions of theestimated on the real (simulated) coverage, over 1000 replicatedsimulations, for different levels of intra-transect coefficient of variation(ITCV, rows) and 4 calculation methods (columns; see Table 1) ofsemi-quantitative survey data

ITCV Method 1 Method 2 Method 3 Method 4

0.1 0.909 0.909 0.985 0.9860.5 0.932 0.936 0.995 0.9951 0.945 0.954 0.996 0.99710 0.963 0.972 0.996 0.997100 0.966 0.973 0.996 0.997Global R2 0.937 0.941 0.993 0.993

High ITCV values correspond to low habitat complexity because of thedominance of one or a few habitat components.


been specifically trained and briefed to correctlyimplement them. The LIT method was applied alongthe 50-m tape. For MSA, ten quadrats were surveyed onboth sides of the tape, for a total of 20 quadrats pertransect. Divers did not exchange information during orafter a survey in order to avoid inconsistencies of theresults among transects. Since the two methods do notcover exactly the same substrate categories, only thecomparable categories have been used in the compar-isons. Two major groups of variables are considered:A—abiotic substrate: No. 1: sand, No. 2: debris, No. 3:soft bottom, No. 4: rock, No. 5: dead coral, No. 6: hardbottom; B—living substrate: No. 7: branching anddigitate coral, No. 8: soft coral, No. 9: encrusting coral,No. 10: branching coral (alone), No. 11: digitate coral(alone), No. 12: massive coral, No. 13: other coral(tabulate, free, fire-corals, foliose). An extra variable wascalculated and added to the analysis, which is No. 14:total coral (both live and dead).

2.5. Statistical analyses of the field survey data

The differences between the two methods (LIT andMSA), in terms of multivariate habitat descriptions, were

Fig. 2. Graphical representation of the five randomly generated virtual lands(Landscape 2), 0.5 (Landscape 3), 1.7 (Landscape 4) and 2.2 (Landscape 5)

investigated by using our balanced sampling design toconduct a 3-factormodel I ANOVA. The 3 factors were (i)the two types of reef (fixed factor: middle, barrier reefs),(ii) the three teams (fixed factor: teams A, B, C) and (iii)the two methods (fixed factor: LIT, MSA). Transectobservations were paired over the last factor because eachteam estimated the habitat components through MSA andLIT in each transect. Multivariate ANOVAs wereperformed through canonical redundancy analysis(RDA). This method allowed us to carry out the analyseson the multivariate response data table (habitat compo-nents) and offered the possibility to test the significance ofeach main factor and interaction term through a MonteCarlo permutation procedure. The main factors and theirinteractions were coded using orthogonal dummy vari-ables (Helmert coding); how to code and test the mainfactors and interactions through canonical analysis aredescribed in Legendre and Anderson (1999). The calcu-lations were performed using the program CANOCO

version 4.5 (ter Braak and Smilauer, 2002). Interactionsbetween factors were first tested to assess the pattern ofvariation; if theywere statistically significant, factors weretested separately in the classes of another factor. In case ofdiverging final results, the probabilities of the kindependent tests were combined using Fisher's methodfor combining the probabilities from independent tests ofsignificance (Sokal and Rohlf, 1995).

To investigate the predictive power of each method interms of assessment of the fish–habitat relationships, weperformed canonical redundancy analysis (RDA) be-tween a matrix Y containing the fish species data pertransect (response variables) and a matrix X containingthe values of 14 environmental variables for each tran-sect (explanatory variables). Biomass was chosen as thebest descriptor of the species fished for human con-sumption. In order to test different aspects of the fishassemblage structure, species biomasses were succes-sively aggregated per mobility patterns (territorial, sed-entary, mobile and very mobile), trophic groups

capes having different coefficients of variation: 0.3 (Landscape 1), 0.4.

Table 4Mean of the squared differences, which estimates the variance due tothe method, between real and estimated surfaces for each virtuallandscape, each one assessed by LIT and MSA

Landscape 1 2 3 4 5

LIT 1622 1562 845 878 808MSA 1551 1499 701 688 626

A smaller mean of squared differences is better.


(piscivores, macro-carnivores, micro-carnivores, zoo-planctivores, other planctivores, macro-herbivores,micro-herbivores, corallivores and detritivores), andsize classes (0–7, 8–15, 16–30, 31–50, 51–80, andN80 cm). For each analysis, a selection of habitat var-iables was first performed using the CANOCO software,which offers a forward selection procedure based uponMonte Carlo tests; non-significant environmental vari-ables were eliminated. A subset of 11 habitat variablesthat had been selected during one of the 6 analyses (No. 1:sand, No. 2: debris, No. 3: soft bottom, No. 4: rock, No. 6:hard bottom, No. 8: soft coral, No. 9: encrusting coral, No.10: branching coral, No. 12: massive coral, No. 13: othercoral, No. 14: total coral) was then used to test thepredictive power of the MSA- and LIT-based habitat datasets on the three groups of fish response variables(mobility patterns, trophic groups, and size classes). TheRDA trace statistic given by the CANOCO software will beused to characterize the success of each analysis. The tracestatistic is equivalent to a coefficient of determination (R2

statistic) since it corresponds to the fraction of the varianceof the fish community explained by the selectedenvironmental variables.

3. Results

3.1. Habitat simulations to validate SQS

The results obtained with the four methods ofcalculation of habitat component coverage percentages(Table 1) were globally acceptable, with R2 rangingfrom 0.909 for the worst (method 1, ITCV=0.1) up to

Table 5Multivariate ANOVA results showing differences in habitat descriptions by

Factors Team A

Trace F P

Reefs×Methods 0.019 0.943 0.422Methods 0.118 5.814 0.010⁎

Between methods for the middle reef onlyBetween methods for the barrier reef only

Reefs×Methods is the interaction between reefs (middle, barrier) and methpermutations: ⁎P≤0.05; ⁎⁎P≤0.01.

0.997 for the best (method 4, ITCV=1, 10 and 100)(Table 3). The coefficient of determination (R2)calculated on all data (aggregation of the five series ofintra-transect coefficients of variation simulations)shows that method 1 was the least accurate (R2 =0.937)whereas method 4 was the most reliable (R2 =0.993).

As shown in Table 3, R2 increased with the intra-transect coefficient of variation (ITCV) for all methods,suggesting that MSA best describes habitats with lowcomplexity (high ITCV values). However, the fourmethods of calculation of habitat component coveragepercentages from semi-quantitative (SQS) survey datadisplayed differences in their R2 levels. Methods 3 and 4produced R2 coefficients above 0.98 for ITCV equal to0.1, whereas the coefficients of determination for meth-ods 1 and 2 were below 91% for that level of ITCV.

Method 4 was used for data integration in the sim-ulations assessing the intrinsic variance of the MSA andLIT estimators (next paragraph), as well as in the fieldsurvey.

3.2. Comparison of the MSA and LIT estimators bysimulation

The five virtual landscapes of decreasing habitatcomplexity are shown in Fig. 2. Relatively homoge-neous landscapes are characterized by a dominant hab-itat component, as exemplified by landscape 5 wherehabitat component 1 represented 75% coverage; moreheterogeneous landscapes are composed of habitatcomponents of similar coverage, as shown by landscape1 where component coverage ranged from 7.1% to18.5% (Table 2, Fig. 2). For both the LIT and MSAmethods, the error made during assessment of the realhabitat coverage (total squared estimation error) de-creased with decreasing landscape complexity (Table 4),indicating lower accuracy for both methods in complexenvironments. However, MSA generated percentagecoverage estimates that where closer to reality than LIT,and this for all values of landscape complexity, as shownin Table 4 where the total squared estimation error

LIT and MSA

Team B Team C

Trace F P Trace F P

0.041 3.200 0.024⁎ 0.028 1.526 0.229⁎ 0.048 2.686 0.054

0.027 1.006 0.4860.361 10.114 0.009⁎⁎

ods (LIT, MSA). F = F-statistics. Probabilities (P) tested using 999

Table 6Multivariate ANOVA results showing the greater variance of LITcompared to MSA for reef habitat description with respect to thehuman factor (Teams) and reef type (Reefs)

Factors MSA LIT

Trace F P Trace F P

Teams×Reefs 0.056 0.799 0.601 0.029 0.497 0.829Reefs 0.051 1.454 0.233 0.114 3.920 0.014⁎

Teams 0.049 0.691 0.670 0.160 2.764 0.034⁎

Team×Reef is the interaction between teams (A, B, C) and reefs(middle, barrier). F = F-statistics. Probabilities (P) tested using 999permutations: ⁎P≤0.05.


was systematically smaller for MSA in all types oflandscapes.

3.3. Comparison of the MSA and LIT estimators duringa field survey

The multivariate ANOVAs revealed that there weresignificant differences between the MSA and LITmethods for habitat description (Table 5). The three-way interaction (Teams×Reefs×Methods) was notsignificant but the Teams×Methods was significant.Separate analyses of the multivariate differences be-tween methods were thus conducted separately by team.For team A, the Reefs×Methods interaction was notsignificant; the 2-way ANOVA revealed a significantdifference between MSA and LIT (P=0.010). For teamB, the Reefs×Methods interaction was significant sothat separate ANOVAs were conducted for each reef; themultivariate difference between methods was notsignificant for the middle reef (P=0.486) but it wassignificant for the barrier reef (P=0.009). For team C,the Reefs×Methods interaction was not significant; the2-way ANOVA revealed a nearly significant differencebetween MSA and LIT (P=0.054). These probabilitieswere combined using Fisher's method (χ2 =−2∑ ln

Table 7Results of linear RDA showing the greater predictive power of fish assembl

MSA

Envir. variables Trace F P

Mobility groups 10, 13, 2, 9 0.655 3.109 0.0Size classes 10, 13, 2, 9 0.667 3.284 0.0Trophic groups 10, 8, 2, 9 0.637 2.866 0.0

Analyses were conducted on fish biomass successively split into 4 groups of mthe 11 environmental variables selected by either MSA or LIT in the prelimina(P≤0.05) are shown in the columns “Envir. variables”; the variables in parenof subsequent significant variables in the models. The identification number“trace” corresponds to the fraction of variance of the species data explainedF-statistics. Probabilities (P) tested using 999 permutations: ⁎P≤0.05; ⁎⁎P

(Pi)=25.912, d.f.=2, k=8) which leads to a highlysignificant combined probability of differences be-tween MSA and LIT (P=0.0011).

The cumulated variance in the habitat description datatables was greater for LIT (total sum of squares=72,482)than for MSA (total sum of squares=70,351). SeparateANOVAs per method (MSA vs. LIT) showed that therewas more variance between reef types in the multivariatedescription of the habitat through LIT (11.4%) than byMSA (5.1%). There was also more variance betweenteams in the habitat estimates made by LIT (16.0%) thanby MSA (4.9%). For both Reefs and Teams, differenceswere not significant for MSA (P=0.233 and P=0.670)whereas they were significant for LIT (P=0.014 andP=0.034) (Table 6). These three approaches concur toshow that LIT produced habitat composition estimatesthat were more variable than MSA.

3.4. Fish–habitat relationships

We will first examine the results of the forwardselection of environmental variables in each canonicalmodel (Table 7, columns “Envir. variables”). For mo-bility groups and size classes, analysis of the relation-ships between species and habitat described by MSArevealed the significant role of four environmentalvariables (branched corals, other corals, debris and en-crusting corals). The same analysis for mobility groupsby habitat described by LIT showed no significant roleof any environmental variable. The analysis with sizeclasses revealed however the structuring role of fourenvironmental variables (soft corals, all corals, hardbottom and sand); two of these (soft coral and hardbottom) were not significant at the P=0.05 level butthey facilitated the entry of subsequent significantvariables in the model. For trophic groups, MSA iden-tified the same 4 environmental variables as in the otherMSA-based analyses, except for “other coral” that was

age structure by MSA compared to LIT

LIT

Envir. variables Trace F P

10⁎⁎ None 0.386 1.197 0.31707⁎⁎ (8), 14, (6), 1 0.537 2.203 0.035⁎

03⁎⁎ (8), (14), 6, 1 0.473 1.703 0.073

obility, 6 size classes and 9 trophic groups. Fish data are explained byry forward selection procedures. The selected variables in each analysistheses have no significant effect (PN0.05) but they facilitated the entrys of environmental variables are given in Materials and methods. Theby the environmental variables; it is equivalent to an R2 statistic. F =≤0.01.


replaced by “encrusting coral”. The same analysis withLIT identified the same 4 variables as for size classes;“soft coral” and “total coral” were not significant at theP=0.05 level, but they facilitated the entry of 2 otherenvironmental variables in the model. All in all, MSAidentified more environmental variables than LIT thatwere significantly related to the fish data.

Using the subset of 11 environmental variablesselected in preliminary analyses by MSA or LIT,canonical analyses showed that the predictive powerof description of the fish assemblage structure based onbiomass by habitat composition was systematicallygreater with MSA compared to LIT (Table 7, columns“trace” and “P”). For mobility groups and size classes,the analyses of the relationships between species andhabitat described by MSA explained respectively 65.5%and 66.7% of the species variance, with highly sig-nificant probabilities (P=0.010 and P=0.007, respec-tively). For trophic groups, MSA explained 63.7% ofthe species variance with a highly significant probability(P=0.003). The same analysis for mobility groups byhabitat described by LIT explained 38.6% of the speciesvariance and a non-significant probability (P=0.317).For size classes, LIT explained 53.7% of the speciesvariance with a significant probability (P=0.035). Fortrophic groups, LIT explained 47.3% of the speciesvariance with a non-significant probability (P=0.073).

Using Fisher's method for combining the probabilitiesof independent tests, we obtained the following globalprobabilities: P=0.00003 for MSA and P=0.02710 forLIT. The more significant canonical analysis resultobtained by MSA indicates that it has more power thanLIT for identifying species–habitat relationships.

4. Discussion

Our goal was to develop an estimator of habitatcomponents having improved characteristics, comparedto the line intercept transect method (LIT): (i) a broaderscale of description of the coral reef habitat; (ii) greaterefficiency, i.e., an estimator having reduced estimationvariance; and (iii) a better assessment of the habitat at ascale compatible with fish community studies. So wedeveloped a medium scale approach (MSA) for habitatassessment and showed how to calculate the habitatcomponents from semi-quantitative data to reach ahighly satisfactory level of accuracy. In our simulations,the coefficients of determination between estimates andreality ranged from 93% to 99%, depending on themethod used for integrating the semi-quantitative datainto quantitative habitat component estimates. Methods1 and 2 present the advantage of requiring no trans-

formation of the semi-quantitative data to a quantitativescale, only normalization, but they showed loweraccuracy for describing the reef habitat. Method 4 waschosen for our study because it consistently producedthe highest coefficients of determination between es-timates and simulated reality. Compared to method 3,method 4 also allows a calculation of intra-transectvariance and a comparison with other methods based onobservations at finer scales. Like method 3, method 4requires a transformation of the semi-quantitative data toa quantitative scale, with the risk of systematic bias.However, if the bias is the same among transects, thedata remain comparable (Craik, 1981).

Using computer-assisted numerical simulations, weshowed that MSA generated less variance than LIT forhabitat description in a complex environment, such as acoral reef ecosystem. A field survey was also used tocompare the two habitat assessment methods. Thesurvey had not been specifically designed for thecomparison of LIT and MSA. For such a purpose, asurvey design based on habitat description within thesame transects by different teams of observers wouldhave been optimal. Since the general objective of thesurvey was to assess the structure of the fish communitytargeted by fishermen and their relationships withhabitat in a very large area (tens of square kilometres),we had to do with the available financing for field timeand human resources; this prevented us from imple-menting an optimal design for the present study. Never-theless, the comparison between teams, habitats, andmethods showed lower estimated variance for MSAcompared to LIT under real field conditions.

We did not attempt to prove the absolute superiority ofMSA compared to LIT, but to show that it was betteradapted to study specific questions, such as the de-scription of the relationships between reef habitat and fishliving in large territories, which include most of thespecies targeted by fisheries. Our study showed that MSAhad higher predictive power than LIT in a study of therelationships between community composition (biomassper species) and habitat characteristics. LIT is extensivelyused for habitat description, to estimate hard coral(Carleton and Done, 1995; Greenstein et al., 1998) orsoft coral coverage (Fabricius, 1997) during underwatersurveys, and also on terrestrial vegetation (Sturges, 1993;Korb et al., 2003). LIT and other closely related methodsare also widely used to describe benthos during reef fishsurveys (e.g. Syms and Jones, 2000; Chateau andWantiez, 2005; McClanahan and Graham, 2005). LITcertainly provides an unbiased description of the substratecover; yet, these descriptions may be ill-adapted to thestudy of broad-ranging species. Our demonstration should


attract the attention of researchers who are using this typeof description for assessing habitat–animals relationshipsfor species with large home ranges, such as largepredatory fish found on reefs (Connell and Kingsford,1998; Gust, 2002) or birds (Call et al., 1992).

Several studies have shown the effect of micro-habitat variables on reef fish structure. Ault and Johnson(1998) demonstrated the significant effect of shelteravailability on species richness and Grigg (1994)identified interstitial spaces as a main contributingfactor for fish abundance. Structural complexity playsan attracting role on reef fishes (Caley and St John,1996). However, in most of these studies, complexitywas assessed at a micro-scale of a few metres (Fig. 3).On the other hand, in several studies, the effect ofmacro-habitats was assessed by relating the type of reef(fringing, intermediary and barrier) to the fish commu-nity composition. Grimaud and Kulbicki (1998) showedthat only 5% to 10% of the species were present in allthree macro-habitats of the New-Caledonian lagoon andthat 45% were limited to a single macro-habitat. A studyof reef fish structure at three scales (regions, reefs, reeftypes) along the eastern coast of the Yucatan Peninsulaof Mexico concluded that the main structuring factorwas the type of reef (habitat type), followed by geo-graphically distinct reefs (Núñez-Lara et al., 2005). Thisstructuring factor has an influence at the communitylevel but also for species, as shown for the density andbiomass of Scarids which varied significantly, in theabsence of fishing effect, along a gradient within midand outer continental shelf positions with local differ-ences between sheltered and exposed sites within eachreef (Gust et al., 2001). Such effects can be detected atmacro-scale (several hundred metres), corresponding tothe landscape approach (Chancerelle, 1996). Betweenmicro- and macro-scales as we define them, there isa meso-scale corresponding to several tens of metres(Fig. 3). At such a scale, variables such as rugosity orreef patch connectivity also play a role, particularly forspecies richness (Luckhurst and Luckhurst, 1978; Ault

Fig. 3. Correspondence between the designations of both scales andtypes of habitat with distances and surfaces involved. The MSA allowsto fill the gap between the landscape approach and the LIT method.

and Johnson, 1998). At that same scale, the MSAestimator has shown in our study that the presence ofcoral could significantly explain the distribution ofbiomass of the fished species (see below). In such acontext of cross-scaling effects, MSA shows a veryinteresting potential for improving the assessment of therelationships between reef fish and their habitat, byimproving the characterization of environmental vari-ables at a scale (grain size, or size of the sampling units)of 500 m2 (two sides×10 quadrats of 25 m2).

In terms of the habitat variables showing an effect onfish structuring, only one environmental variable (No. 8,soft coral, with a significant effect with MSA only on“trophic groups”) was identified as a structuring factorby both methods. This result is somewhat unprecedent-ed and may be explained by both the inter-transectvariability and particularly the very high complexity ofthe structuring processes of reef fish stocks. We assumedin our study that the habitat variables were a majorstructuring factor of consumed fish populations, asshown by several authors (e.g. Jennings et al., 1996), inparticular if we compare it with fishing effects (Clua,2004). Other structuring factors are known to havesignificant effects, such as recruitment (Sale, 1991;Hixon and Webster, 2002), inter-specific predation(Caley, 1993), fishing (Russ and Alcala, 1998; Jenningsand Kaiser, 1998) or temporal variability (Galzin, 1987;Friedlander and Parrish, 1998b; Thompson and Map-stone, 2002). Considering the “soft coral” variable itself,its structuring role identified in our study should beinterpreted with caution, since other authors have notbeen able to prove any direct effect on fish assemblagestructure during specific experiments (Syms and Jones,2001). These authors suggest that soft corals may affectfish assemblages indirectly, by occupying space thatwould otherwise be covered by hard corals. None of theother variables highlighted by MSA (i.e., branchingcoral, other coral, encrusting coral or debris) showed upin the LIT variable selection analyses. As far as weknow, the literature does not mention these variables asusual structuring variables of fish populations, exceptfor “branching coral” in a study conducted in Hawaii(Friedlander et al., 2003). On the other hand, manyauthors have already shown the structuring role ofenvironmental variables such as “live coral” (Bell andGalzin, 1984; Legendre et al., 1997) or “hard bottom”and “sand” (Labrosse, 2000; Clua, 2004), as revealed bythe LIT method in our study on biomass by size classesand trophic groups.

Most of the reef species targeted by fishing are mo-bile and have extensive home ranges, and sometimesshy behaviour in the presence of divers. The distance


sampling method (Buckland et al., 2001) allows usersto better estimate, during underwater visual censuses(UVC), that part of the fish community which would beleft out by other visual census methods, like transectswith predetermined width or fixed counting points. Therelationships between these species and their habitatmust be studied at a suitable scale. That constraint ismet by the MSA method, which was shown in thisstudy to be superior to LIT. In the future, it would beinteresting to compare, in the same way, the MSA andRRI estimators (Long et al., 2004), using bothcomputer-generated and field data. It is likely that asimilar result would be obtained, since the RRI methodsurveys at a scale smaller than that of MSA andcomparable to that of LIT. The MSA, which targets theintermediate scale (b1000 m2) between the scale whereLIT is appropriate (b100 m2) and that of the landscapeapproach (N1000 m2), enriches the toolbox of methodsavailable for reef habitat description at different scales(Fig. 3).

Acknowledgements

This work was funded, and field work was im-plemented, by the Northern Province of New Caledonia,Falconbridge Ltd., the Secretariat of the PacificCommunity, and IRD (Institut de Recherche pour leDéveloppement). We are grateful to the crew of theNorthern Province Fisheries Department boat «Max»and the divers G. Mou-Tham and S. Sauni for invaluableassistance in the field, and to Yves-Marie Bozec as amember of the team that initiated the MSAmethod. Thisresearch was also supported by NSERC grant no.OGP0007738 to P. Legendre.

References

Ault, T.R., Johnson, C.R., 1998. Spatial variation in fish speciesrichness on coral reefs: habitat fragmentation and stochastic struc-turing processes. Oikos 82, 354–364.

Bell, J.D., Galzin, R., 1984. Influence of live coral cover on coral reeffish communities. Mar. Ecol. Prog. Ser. 15, 265–274.

Bellehumeur, C., Legendre, P., Marcotte, D., 1997. Variance andspatial scales in a tropical rain forest: changing the size of samplingunits. Plant Ecol. (Dordr.) 130, 89–98.

Beukers, J.S., Jones, G.P., 1998. Habitat complexity modifies theimpact of piscivores on a coral reef fish population. Oecologia 114,50–59.

Buckland, S.T., Anderson, D.R., Burnham, K.P., Laake, J.L.,Borchers, D.L., Thomas, L., 2001. Introduction to distance sam-pling. Estimating Abundance of Biological Populations. OxfordUniversity Press, Oxford.

Caley, M.J., 1993. Predation, recruitment and the dynamics ofcommunities of coral-reef fishes. Mar. Biol. 117, 33–43.

Caley, M.J., St John, J., 1996. Refuge availability structures assem-blages of tropical reef fishes. J. Anim. Ecol. 65, 414–428.

Call, D.R., Guttierez, R.J., Verner, J., 1992. Foraging habitat andhome-range characteristics of California spotted owls in the SierraNevada. Condor 94, 880–888.

Carleton, J.H., Done, T.J., 1995. Quantitative video sampling of coralreef benthos: large-scale application. Coral Reefs (HistoricalArchive) 14, 35–46.

Chancerelle, Y., 1996. Caractérisation des paysages récifaux sous-marins de l'île de Mooréa (Polynésie Française). Thèse dedoctorat, Université de la Polynésie Française.

Chateau, O., Wantiez, L., 2005. Comparaison de la structure descommunautés de poissons coralliens entre une réserve marine etdeux zones proches non protégées dans le parc du lagon sud de laNouvelle-Calédonie. Cybium 29 (2), 159–174.

Clua, E., 2004. Influence relative des facteurs écologiques et de lapêche sur la structuration des stocks de poissons récifaux dans sixpêcheries du Royaume des Tonga (Pacifique Sud). Thèse dedoctorat, École Pratique des Hautes Études, La Sorbonne, Paris.

Craik, G., 1981. Underwater survey of coral trout Plectropomusleopardus (Serranidae) populations in the Capricornia section ofthe Great Barrier Reef Marine Park. Proc. 4th Internat. Coral ReefSymp., Manila, Philippines, vol. 1, pp. 53–58.

Connell, S.D., Kingsford, M.J., 1998. Spatial, temporal and habitat-related variation in the abundance of large predatory fish at OneTree Reef Australia. Coral Reefs 17, 49–57.

English, S., Wilkinson, C., Baker, V., 1997. Survey Manual forTropical Marine Resources, 2nd edition. AIMS, Townsville.

Fabricius, K.E., 1997. Soft coral abundance in the central Great BarrierReef: effects of Acanthaster planci, space availability and aspectsof the physical environment. Coral Reefs 16, 159–167.

Friedlander, A.M., Parrish, J.D., 1998a. Habitat characteristicsaffecting fish assemblages on a Hawaiian coral reef. J. Exp. Mar.Biol. Ecol. 224, 1–30.

Friedlander, A.M., Parrish, J.D., 1998b. Temporal dynamics of fishcommunities on an exposed shoreline in Hawaii. Environ. Biol.Fisches 53, 1–18.

Friedlander, A.M., Brown, E.K., Jokiel, P.L., Smith, W.R., Rodgers,K.S., 2003. Effects of habitat, wave exposure, and marine pro-tected area status on coral reef fish assemblages in the Hawaiianarchipelago. Coral Reefs 22, 291–305.

Galzin, R., 1987. Structure of fish communities of French Polynesia.II: temporal scales. Mar. Ecol. Prog. Ser. 41, 137–145.

Galzin, R., Harmelin-Vivien, M.L., 2000. Ecologie des poissons desrécifs coralliens. In: Pichon et, M., Duffour, V. (Eds.), Les récifscoralliens, Oceanis, Documents océanographiques, secondepartie. Aspects Biologiques, 26 (3), pp. 465–495.

González-Gándara, C., Membrillo-Venegas, N., Núñez-Lara, E.,Arias-González, J.E., 1999. The relationship between fish andreefscapes in the Alacranes Reef Yucatan, Mexico. A preliminarytrophic functioning analysis. Vie Milieu 49, 275–286.

Greenstein, B.J., Curran, H.A., Pandolfi, J.M., 1998. Shiftingecological baselines and the demise of Acropora cervicornis inthe western North Atlantic and Caribbean Province: a Pleistoceneperspective. Coral Reefs 17, 249–261.

Grigg, R.W., 1994. Effects of sewage discharge fishing pressure andhabitat complexity on coral ecosystems and reef fishes in Hawaii.Mar. Ecol. Prog. Ser. 103, 25–34.

Grimaud, J., Kulbicki, M., 1998. Influence de la distance a l'océansur les peuplements ichtyologiques des récifs frangeants deNouvelle-Calédonie. C. R. Hebd. Séances Acad. Sci., Sér. 3, Sci.Vie, vol. 321, pp. 923–931.


Gust, N., 2002. Scarid biomass on the northern Great Barrier Reef: theinfluence of exposure, depth and substrata. Environ. Biol. Fisches64, 353–366.

Gust, N., Choat, H.J., McCormick, M.I., 2001. Spatial variability inreef fish distribution, abundance, size and biomass: a multi-scaleanalysis. Mar. Ecol. Prog. Ser. 214, 237–251.

Hixon, M.A., Webster, M.S., 2002. Density dependence in reef fishpopulations. In: Sale, P.F. (Ed.), Coral reef fishes — Dynamicsand diversity in a complex system. Academic Press, London,pp. 303–325.

Hurlbert, S.H., 1984. Pseudoreplication and the design of ecologicalfield experiments. Ecol. Monogr. 54, 187–211.

Jennings, S., Kaiser, M.J., 1998. The effects of fishing on marineecosystems. Adv. Mar. Biol. 34, 201–352.

Jennings, S., Boulle, D.P., Polunin, N.V.C., 1996. Habitat correlates ofthe distribution and biomass of Seychelles' reef fishes. Environ.Biol. Fisches 46, 15–25.

Jones, G.P., Syms, C., 1998. Disturbance habitat structure and theecology of fishes on coral reefs. Aust. J. Ecol. 23, 287–297.

Kolasa, J., Rollo, C.D., 1991. Introduction: the heterogeneity ofheterogeneity: a glossary. In: Kolasa, J., Picket, S.T.A. (Eds.),Ecological Heterogeneity. Springer-Verlag, New York, pp. 1–23.

Korb, J.E., Covington, W.W., Fulé, P.Z., 2003. Sampling techniquesinfluence understory plant trajectories after restoration: an examplefrom ponderosa pine restoration. Restor. Ecol. 11, 504–515.

Kulbicki, M., Sarramegna, S., 1999. Comparison of densityestimates derived from strip transect and distance sampling forunderwater visual censuses: a case study of Chaetodontidae andPomacanthidae. Aquat. Living Resour. 12, 315–325.

Kulbicki, M., Letourneur, Y., Labrosse, P., 2000. Fish stockassessment of the northern New Caledonian lagoons: 2-stocks oflagoon bottom and reef-associated fishes. Aquat. Living Resour.13, 77–90.

Labrosse, P., 2000. Structure des peuplements des poissons récifauxcommerciaux de la Province Nord de Nouvelle-Calédonie.Mémoire de Diplôme de l'École Pratique des Hautes Études, LaSorbonne, Paris.

Labrosse, P., Kulbicki, M., Ferraris, J., 2003. Underwater visual fishcensus surveys. Reef Ressources Assessment Tools Collection.Secretariat of the Pacific Community (ed), Noumea, NewCaledonia.

Legendre, P., Anderson, M.J., 1999. Distance-based redundancyanalysis: testing multispecies responses in multifactorial ecologicalexperiments. Ecol. Monogr. 69, 1–24.

Legendre, P., Galzin, R., Harmelin-Vivien, M.L., 1997. Relatingbehavior to habitat: solutions to the fourth-corner problem.Ecology 78, 547–562.

Letourneur, Y., Kulbicki, M., Labrosse, P., 1998. Length–weightrelationships of fish from coral reefs and lagoons of NewCaledonia, southwestern Pacific Ocean: an update. NagaICLARM Q, vol. 21, pp. 39–46.

Long, B.G., Andrews, G., Suharsono, Y.G.W., 2004. Samplingaccuracy of reef resource inventory technique. Coral Reefs 23,378–385.

Luckhurst, B.E., Luckhurst, K., 1978. Analysis of the influence of thesubstrate variables on coral reef fish communities. Mar. Biol. 49,317–323.

McClanahan, T.R., 1994. Kenyan coral reef lagoon fish: effects offishing, substrate complexity and sea urchins. Coral Reefs 13,231–241.

McClanahan, T.R., Graham, N.A.J., 2005. Recovery trajectories ofcoral reef fish assemblages within Kenyan marine protected areas.Mar. Ecol. Prog. Ser. 294, 241–248.

Mikulski, P.W., 1982. Efficiency asymptotic relative (ARE) ofestimators. In: Kotz, S., Johnson, N.L. (Eds.), Encyclopedia ofStatistical Sciences, vol. 2. Wiley, New York, pp. 468–469.

Núñez-Lara, E., Arias-González, J.E., Legendre, P., 2005. Spatialpatterns of Yucatan reef fish communities: testing models using amulti-scale survey design. J. Exp. Mar. Biol. Ecol. 324, 157–169.

Russ, G.R., Alcala, A.C., 1998. Natural fishing experiments in marinereserves 1983–1993: community and trophic responses. CoralReefs 17, 383–397.

Sale, P., 1991. The Ecology of Fishes on Coral Reefs. Academic Press,San Diego.

Sokal, R.R., Rohlf, F.J., 1995. Biometry — The Principles andPractice of Statistics in Biological Research, 3rd edition. W.H.Freeman, New York.

Sturges, D., 1993. Soil–water and vegetation dynamics through 20years after big sagebrush control. J. Range Manag. 46, 161–169.

Syms, C., Jones, G.P., 2000. Disturbance, habitat structure, and thedynamics of a coral-reef fish community. Ecology 81 (10),2714–2729.

Syms, C., Jones, G.P., 2001. Soft corals exert no direct effects on coralreef fish assemblages. Oecologia 127, 560–571.

ter Braak, C.J.F., Smilauer, P., 2002. CANOCO Reference Manual andCanoDraw for Windows User's Guide: Software for CanonicalCommunity Ordination (version 4.5). Microcomputer Power,Ithaca, New York.

Thompson, A.A., Mapstone, B.D., 2002. Intra- versus inter-annualvariation in counts of reef fishes and interpretations of long-termmonitoring studies. Mar. Ecol. Prog. Ser. 232, 247–257.